A high-fidelity stitching method and system for large field-of-view hyperspectral microscopic images

Abstract

This invention discloses a high-fidelity stitching method and system for large-field hyperspectral microscopic images, belonging to the field of hyperspectral microscopic imaging and digital pathology image processing technology. The invention performs system-level physical calibration offline, establishing wavelength-wise spatial mapping relationships, a full-spectrum energy attenuation model, and a unified absolute physical coordinate grid. In the online stage, wavelength-wise radiometric correction and absolute coordinate restoration are performed on the multi-field hyperspectral scanning data to complete the initial global arrangement of each field of view. A fusion cost function is constructed based on the spatial-spectral gradient continuity criterion, and the optimal stitching boundary is solved through optimization methods. Adaptive fusion of overlapping areas is performed using a single-field pixel optimization writing method. Finally, the processed field-view data are integrated into a panoramic hyperspectral data cube. This invention reduces wavelength-related spatial mismatch, improves edge spectral drift, and avoids damage to microstructures caused by non-physical elastic deformation, making it suitable for high-precision microscopic analysis scenarios such as digital pathology.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral microscopy and digital pathology image processing technology, and in particular to a high-fidelity stitching method and system for large-field hyperspectral microscopic images. Background Technology

[0002] Hyperspectral microscopy can simultaneously acquire spatial morphological information and continuous spectral information of target samples, and can provide biochemical differences in cells, blood vessels, tissue stroma and lesion areas. It has important application value in digital pathology, tumor-assisted diagnosis and tissue optical characterization.

[0003] In high-magnification microscopy, while high-magnification objectives offer high spatial resolution, their limited field of view in a single acquisition makes it difficult to cover large tissue sections. Low-magnification objectives, on the other hand, offer a wider field of view but lack sufficient spatial resolution, failing to identify cellular and microvascular pathological features. Therefore, the industry commonly employs a moving stage to scan and acquire multiple overlapping local images per field of view, then uses stitching technology to construct a panoramic hyperspectral data cube, thus achieving both a large field of view and high resolution.

[0004] However, existing hyperspectral microscopy stitching methods still have many shortcomings in engineering applications: 1) Wavelength-dependent spatial distortion and spectral distortion: High-magnification microscope objectives suffer from higher-order radial distortion, tangential distortion, and wavelength-dependent image shift caused by dispersion. Pixel shifts are inconsistent across different wavelength bands, and geometric correction of a single reference band cannot guarantee the consistency of spatial position across the entire band. Simultaneously, inhomogeneities in the transmitted illumination system, lens vignetting, and detector response differences lead to baseline drift and peak attenuation in the field-of-view spectrum, affecting subsequent quantitative analysis.

[0005] 2) Traditional stitching methods disrupt microscopic physical scale and spectral features: To improve visual continuity, existing techniques often employ non-rigid registration, elastic deformation, or local stretching correction, which alters pathologically significant microstructures such as cell area and blood vessel diameter. Furthermore, regarding spectral inconsistencies in overlapping areas, commonly used methods such as mean matching and statistical smoothing struggle to distinguish between physical attenuation and true pathological spectral differences, easily erasing crucial spectral information such as weak edge peaks and narrow-band absorption features, thus weakening the quantitative analysis value of hyperspectral images. Summary of the Invention

[0006] The purpose of this invention is to provide a high-fidelity stitching method and system for large-field hyperspectral microscopic images, overcoming the shortcomings of existing image stitching technologies such as elastic stretching altering physical morphology and statistical filtering erasing true pathological spectra. By establishing wavelength-independent spatial-physical inverse mapping relationships and a full-spectrum energy attenuation model in the offline stage, and combining absolute coordinate restoration and spatial-spectral gradient continuity optimization in the overlapping area in the online scanning stage, large-field stitching of multi-field hyperspectral microscopic images can be achieved.

[0007] To achieve the above objectives, the present invention provides a high-fidelity stitching method for large-field hyperspectral microscopic images, comprising the following steps: S1. System-level physical calibration: Establishing wavelength-dependent spatial mapping relationships, full-spectrum energy attenuation models, and a unified absolute physical coordinate grid for panoramic stitching in an offline state; S2. Online acquisition and physical correction: Perform multi-field hyperspectral scanning on the target sample to obtain multiple overlapping local hyperspectral data cubes; perform physical correction on each local data cube and map it to the unified absolute physical coordinate grid to complete the initial global arrangement of each field of view; S3. Adaptive fusion of overlapping regions: Based on the spatial-spectral gradient continuity criterion, adaptive fusion processing is performed on the overlapping regions of adjacent fields of view; S4. Panoramic Data Construction: Integrate all fused field-of-view data into a unified absolute physical coordinate grid to generate a panoramic hyperspectral data cube.

[0008] Preferably, S1 specifically includes the following steps: S11. Wavelength-by-wavelength spatial distortion calibration: Using a calibration target with known absolute physical dimensions, calibration images are acquired at multiple consecutive wavelengths. For each wavelength, a mapping relationship or its inverse mapping relationship from the acquired pixel coordinate system to the physical coordinate system is established, forming a set of inverse mapping data indexed by wavelength. S12. Full-spectrum energy attenuation model calibration: Collect full-band dark field and bright field data, and calibrate the model for each wavelength. Construct the corresponding dark field image and bright field image The energy attenuation distribution function at this wavelength is obtained by smoothing and fitting the bright field image. By combining the attenuation distribution functions of all wavelengths, a full-spectrum energy attenuation model is formed; S13. Establishment of a unified absolute physical coordinate grid: Define the unified absolute physical coordinate grid based on the physical size of the calibrated target, the pixel size of the detector, the imaging magnification, and the imaging ratio of the system.

[0009] Preferably, in S11, the calibration target is a high-precision grid target or a lattice target; the implementation of the mapping relationship or inverse mapping relationship includes radial distortion model, radial and tangential joint distortion model, polynomial geometric model, spline model and lookup table.

[0010] Preferably, in S12, the smoothing fitting method for the bright field image includes two-dimensional polynomial fitting, low-rank surface fitting, and spline surface fitting; the full-spectrum energy attenuation model is used to perform radiometric correction on the original hyperspectral data, and the expression of the full-spectrum energy attenuation model is: .

[0011] Preferably, S2 specifically includes the following: S21. Multi-field hyperspectral scanning acquisition: A hyperspectral microscopy imaging device and a motorized stage are used to scan the sample field by field, obtaining multiple overlapping hyperspectral local data cubes, and recording the stage position information corresponding to each field of view; the overlap rate between adjacent fields of view in the scanning direction is 10% to 30%; the scanning path includes line scanning, serpentine scanning, and block scanning; k The raw hyperspectral data obtained from each field of view are: ; S22. Radiometric Correction: Using the dark field data and full-spectrum energy attenuation model from S12, the corrected hyperspectral data is obtained. The correction method is wavelength-by-wavelength radiative compensation, calculated using the following formula: ; in, For raw hyperspectral data, This is the dark field image corresponding to wavelength λ. wavelength λ The corresponding full-spectrum energy decay model; A small positive constant set to prevent the denominator from being too small; S23. Wavelength-by-wavelength absolute coordinate restoration: For each field of view after radiometric correction, the corresponding inverse mapping data is called at each wavelength to reverse map the pixels from the acquisition coordinate system to the unified absolute physical coordinate grid, and the field of view is located to the initial position of the panoramic coordinate system in combination with the stage position information; the inverse mapping method includes bilinear interpolation and bicubic interpolation, which are used to obtain the pixel values ​​at non-integer coordinates. S24. Establishment of panoramic frame to be fused and identification of overlapping areas: Based on the initial position, field size and preset overlap relationship of each field of view determined in S23, establish the panoramic data frame to be fused and identify the overlapping areas between adjacent fields of view.

[0012] Preferably, S3 specifically includes the following: S31. Overlapping Area Determination: Extract any two adjacent fields of view under a unified absolute physical coordinate system. A and B The overlapping region Ω; S32. Joint gradient feature construction: Calculate the features of two adjacent views within the overlapping region Ω. A and B The spatial gradient and spectral gradient are combined to form a joint spatial-spectral gradient feature; S33. Constructing the fusion cost function and solving for the optimal splicing boundary: Constructing the fusion cost function with the goal of minimizing the gradient abrupt change at the splicing seam, and solving for the optimal splicing boundary path through optimization methods; S34, Single-view pixel optimization writing: Divide the overlapping area into two sub-regions according to the optimal stitching boundary path, and each sub-region retains the pixels from the corresponding single view as the final output.

[0013] Preferably, in S32, for any position The spatial gradient is defined as: ; The spectral gradient is defined as: ; The joint spatial-spectral gradient characteristics are: ; in, G xy ( x , y , λ () is the spatial gradient vector, describing the image at position. Local spatial structure changes; Image intensity I exist x The partial derivative in the direction reflects the rate of change in the horizontal direction; Image intensity I exist y The partial derivative in the direction reflects the rate of change in the vertical direction; G λ ( x , y , λ The spectral gradient represents the rate of change of light intensity in a hyperspectral image along the wavelength direction. Image intensity I For wavelength λ The partial derivatives; These are weighting coefficients used to balance the dimensional differences between spatial and spectral gradients.

[0014] Preferably, in S33, the fusion cost function is: ; in: These are non-negative weighting coefficients; and Representing the field of view and vision In position Joint gradient at; and Representing the field of view and vision In position Image intensity value at; This is a confidence penalty term used to characterize the local noise level, edge signal-to-noise ratio, and penalty for distance from the view boundary; The boundary smoothing term is used to constrain the continuity of the spliced ​​boundary path and avoid excessively jagged boundaries; the optimization methods include dynamic programming, graph cut, shortest path search and graph optimization.

[0015] Preferably, in S34, the specific operation is as follows: dividing the overlapping region into areas belonging to the field of view. A and vision B Two sub-regions; for any overlapping region position p The final output pixel value is represented as: .

[0016] The present invention also provides a high-fidelity stitching system for large-field hyperspectral microscopic images, comprising: The system calibration module is used to establish wavelength-dependent spatial mapping relationships, full-spectrum energy attenuation models, and a unified absolute physical coordinate grid for panoramic stitching in an offline state. The online acquisition and correction module is used to perform multi-field hyperspectral scanning of the target sample to obtain multiple overlapping local hyperspectral data cubes. Each local data cube is physically corrected and mapped to the unified absolute physical coordinate grid to complete the initial global arrangement of each field of view. The overlapping region fusion module is used to adaptively fuse overlapping regions of adjacent fields of view based on the spatial-spectral gradient continuity criterion. The panorama construction module is used to integrate all fused field-of-view data into a unified absolute physical coordinate grid to generate a panorama hyperspectral data cube.

[0017] Therefore, the high-fidelity stitching method and system for large-field hyperspectral microscopic images described above have the following beneficial effects: 1) Reduce wavelength-dependent spatial mismatch: By establishing spatial inverse mapping relationships independently for each wavelength, rather than using a unified geometric correction for a single reference band, it is beneficial to reduce cross-band spatial misalignment caused by dispersion and higher-order optical distortion.

[0018] 2) Improve the edge spectral drift problem: Radiometric correction is performed by using dark field, bright field and full spectrum energy attenuation models, which helps to reduce the impact of uneven lighting, dark corners and device response differences on the edge spectrum.

[0019] 3) Avoid damage to the microstructure scale by non-physical elastic deformation: Spatial correction is completed based on physical calibration and absolute coordinate restoration, without relying on free elastic stretching of pathological tissue structures, which is conducive to maintaining the physical consistency of micromorphology such as cell, blood vessel and tissue boundary.

[0020] 4) Reduce the smoothing of the true spectral shape by statistical equalization: Pixel selection in the overlapping area is performed by using the spatial-spectral gradient continuity criterion, rather than global statistical equalization of the entire overlapping area. This is beneficial for preserving local weak peaks and differences in the true pathologically relevant spectrum.

[0021] 5) Suitable for constructing panoramic hyperspectral microscopy images with large field of view: Unifying multi-field high-resolution hyperspectral microscopy images into the same absolute physical coordinate system is beneficial for generating panoramic hyperspectral data cubes suitable for subsequent quantitative analysis.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the hyperspectral microscopy splicing system described in this invention; Figure 2 This is a flowchart illustrating the overall process of the hyperspectral microscopy stitching method described in this invention. Figure 3 This is a schematic diagram illustrating the establishment of the wavelength-by-wavelength spatial distortion calibration and inverse mapping relationship in this invention; Figure 4 This is a schematic diagram illustrating the calibration of the full-spectrum energy decay model and the correction of radiometric measurements in this invention; Figure 5 This is a schematic diagram of multi-field scanning and unified absolute physical coordinate restoration in this invention; Figure 6 This is a schematic diagram of overlapping region fusion based on spatial-spectral gradient continuity in this invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] Example 1 This embodiment provides a high-fidelity stitching method for large-field hyperspectral microscopic images, applicable to the large-field panoramic hyperspectral data cube reconstruction after multi-field scanning of fixed tissue sections, pathological sections, or other thin-layer biological samples using a hyperspectral microscopy imaging system. The hyperspectral microscopy stitching system used in this embodiment is as follows: Figure 1 As shown, it includes a microscope body, an illumination module, a motorized stage, a hyperspectral acquisition module, an image sensor, an image processing unit, and a display / storage unit. The hyperspectral acquisition module is used to acquire image data of the sample at multiple consecutive wavelengths. The motorized stage is used to drive the sample to perform field-by-field scanning in the XY plane. The image processing unit is used to execute the entire algorithm process of system-level physical calibration, online acquisition correction, adaptive fusion of overlapping areas, and panoramic data construction. The display / storage unit is used to display and store the final hyperspectral data.

[0027] Figure 2 This is a flowchart illustrating the overall process of the hyperspectral microscopy stitching method described in this embodiment of the invention. The following is a summary of the process. Figure 2 The accompanying drawings and subsequent figures provide a detailed description of a specific implementation process of this method.

[0028] S1, System-level physical calibration.

[0029] After the system is first installed, regularly maintained, or the optical components are replaced, a system-level physical calibration is performed offline to establish wavelength-dependent spatial mapping relationships, a full-spectrum energy attenuation model, and a unified absolute physical coordinate grid for panoramic stitching. The purpose is to establish physical prior knowledge for subsequent online scanning.

[0030] S11. Wavelength-by-wavelength spatial distortion calibration: like Figure 3As shown, a high-precision grid target (or dot matrix target) with known absolute physical dimensions is placed on the stage. Calibration images are acquired at multiple consecutive wavelengths using the hyperspectral acquisition module. For each wavelength, the image coordinates of feature points on the target are extracted, and combined with the actual physical coordinates of the corresponding feature points, a pixel coordinate system for that wavelength is established. To the physical coordinate system The mapping relationship or its inverse mapping relationship is used to form a wavelength-indexed inverse mapping data set for subsequent online correction. Specifically, it is expressed as follows: Mapping relationship: ; Inverse mapping relationship: ; The implementation methods of mapping or inverse mapping relationships include, but are not limited to: higher-order radial distortion models, radial and tangential joint distortion models, polynomial geometric models, spline models, and lookup table forms. Among them, a preferred approach is to establish a set of independent distortion parameters or independent inverse mapping lookup tables for each wavelength λ to characterize the spatial distortion characteristics of the microscopic system at that wavelength.

[0031] S12. Calibration of the full-spectrum energy decay model: like Figure 4 As shown, full-band dark field and bright field data were acquired. Specifically, using a blank glass slide or a uniformly transmitted sample, dark field data (used to characterize detector dark current and background noise) was acquired with the illumination source turned off across the entire band, while bright field data (used to characterize light source illumination distribution, lens vignetting, and detector response non-uniformity) was acquired with the illumination source turned on.

[0032] For each wavelength Construct the corresponding dark field image and bright field image The energy attenuation distribution function at this wavelength is obtained by smoothing and fitting the bright field image. Smooth fitting can be achieved using methods such as two-dimensional polynomial fitting, low-rank surface fitting, or spline surface fitting. Combining the attenuation distribution functions corresponding to all wavelengths forms a full-spectrum energy attenuation model. This model is used for subsequent radiometric correction of the original hyperspectral data.

[0033] S13. Establishment of a unified absolute physical coordinate grid: Based on the known physical dimensions of the calibrated target, the detector pixel size, the magnification of the microscopic imaging, and the system imaging ratio, a unified absolute physical coordinate grid is defined for panoramic stitching. This unified absolute physical coordinate grid can be a two-dimensional grid with a fixed physical sampling interval (e.g., in micrometers), serving as a global reference frame for all subsequent field-of-view and band mappings. All subsequently acquired field-of-view data will be uniformly mapped to this physical coordinate grid, ensuring physical scale consistency across the entire field of view and band.

[0034] S2. Online data acquisition and physical correction.

[0035] After completing the system-level physical calibration, multi-field hyperspectral scanning acquisition was performed on the tissue slices to be tested, and physical correction and coordinate mapping were performed on each local data cube.

[0036] S21. Multi-field hyperspectral scanning acquisition: The tissue slices to be tested were fixed on the stage. Using a hyperspectral microscopy imaging device and a motorized stage, the target sample was scanned and acquired field-by-field under a high-magnification objective lens, resulting in multiple overlapping hyperspectral local data cubes. The stage position information corresponding to each field of view was recorded. The overlap rate of adjacent fields of view in the scanning direction was set to 10% to 30% to balance stitching stability and scanning efficiency. The scanning path can employ line scanning, serpentine scanning, or block scanning. Let the first... k The raw hyperspectral data obtained from each field of view acquisition are .

[0037] S22, Radiation Measurement Correction: For each field of view's raw hyperspectral data, the dark field data and full-spectrum energy attenuation model obtained in S12 are used for correction. The correction method is as follows: The formula for wavelength-by-wavelength radiative compensation is as follows: .

[0038] in, For raw hyperspectral data, This is the dark field image corresponding to wavelength λ. wavelength λ The corresponding full-spectrum energy decay model; A positive small constant is set to prevent the denominator from being too small.

[0039] After the above radiometric correction, the problems of spectral baseline drift and peak attenuation at the edge of the field of view caused by uneven illumination, lens vignetting and detector response differences can be effectively eliminated.

[0040] S23. Wavelength-by-wavelength absolute coordinate restoration: like Figure 5As shown, for each field of view after radiometric correction, at each wavelength λ, the wavelength-indexed inverse mapping data established in S11 is called to reverse map the pixels from the acquisition coordinate system to the unified absolute physical coordinate grid, thus obtaining the hyperspectral data of that field of view in the absolute physical coordinate system. During mapping, bilinear interpolation or bicubic interpolation can be used to obtain pixel values ​​at non-integer coordinates. Simultaneously, by combining the position information recorded by the stage during acquisition, each field of view is positioned to its initial position in the panoramic coordinate system. Thus, before entering the fusion process, each field of view is unified to the same physical scale and global reference coordinate system.

[0041] S24. Establishment of the panoramic frame to be fused and identification of overlapping areas: Based on the initial positions of each field of view, the physical dimensions of each individual field of view, and the preset overlap relationships determined in S23, an initial global arrangement is performed on all corrected fields of view to form a panoramic data framework to be fused. Under a unified absolute physical coordinate system, the overlapping region between any two adjacent fields of view is identified, where... A , B This refers to the numbering of two adjacent fields of view that have an overlapping relationship.

[0042] S3, Adaptive fusion of overlapping areas.

[0043] like Figure 6 As shown, based on the spatial-spectral gradient continuity criterion, adaptive fusion processing is performed on the overlapping regions of adjacent fields of view to avoid the destruction of physical morphology and spectral features by traditional fusion methods.

[0044] S31. Determination of overlapping area: Under a unified absolute physical coordinate system, for any two adjacent views that overlap, A and B Extract the overlapping region Ω.

[0045] S32. Construction of joint gradient features: Calculate the two adjacent fields of view within the overlapping region Ω. A and B Spatial gradient and spectral gradient.

[0046] For any position The spatial gradient is defined as: ; The spectral gradient is defined as: ; Combining spatial gradients and spectral gradients to form a joint spatial-spectral gradient feature: ; in, G xy (x , y , λ () is the spatial gradient vector, describing the image at position. Local spatial structure changes; Image intensity I exist x The partial derivative in the direction reflects the rate of change in the horizontal direction; Image intensity I exist y The partial derivative in the direction reflects the rate of change in the vertical direction; G λ ( x , y , λ The spectral gradient represents the rate of change of light intensity in a hyperspectral image along the wavelength direction. Image intensity I For wavelength λ The partial derivatives; These are weighting coefficients used to balance the dimensional differences between spatial and spectral gradients.

[0047] S33. Construct the fusion cost function and solve for the optimal stitching boundary: A fusion cost function is constructed with the objective of minimizing gradient abrupt changes across the splicing seam, and the optimal splicing boundary path is obtained by solving the optimization method. Specifically, the fusion cost function is: ; in: These are non-negative weighting coefficients; and Representing the field of view and vision In position Joint gradient at; and Representing the field of view and vision In position Image intensity value at; This is a confidence penalty term used to characterize the local noise level, edge signal-to-noise ratio, and penalty for distance from the view boundary; The boundary smoothing term is used to constrain the continuity of the splicing boundary path and avoid excessively jagged boundaries. The optimal splicing boundary path is obtained by solving the problem using optimization methods such as dynamic programming, graph cut, shortest path search, or graph optimization.

[0048] S34, Single-view pixel preferred writing: Based on the obtained optimal stitching boundary path, the overlapping region Ω is divided into sections belonging to the field of view. A and vision BTwo sub-regions. For any overlapping region position p The final output pixel value is represented as: ; That is, each sub-region strictly retains the corrected pixels from the corresponding single field of view as the final output, rather than performing a weighted average over the entire overlapping area. This embodiment also provides another optional implementation method, in which a narrow transition band of preset width is set on both sides of the optimal boundary for small-scale smoothing, but the width of the transition band is strictly limited to avoid spectral distortion caused by large-scale statistical equalization.

[0049] S4, Panoramic Data Construction.

[0050] The results of processing all fields of view through steps S2 and S3 are written into a unified absolute physical coordinate grid to generate a panoramic hyperspectral data cube. This data cube maintains the continuity of spatial structure and the fidelity of spectral features, and can be further used for subsequent processing tasks such as tissue classification, lesion detection, peak analysis, quantitative measurement, digital pathology slide reading, and machine learning modeling.

[0051] Therefore, this invention employs the aforementioned high-fidelity stitching method and system for large-field hyperspectral microscopic images. By combining "wavelength-dependent physical calibration" with "spatial-spectral gradient continuity optimization," it achieves the following beneficial effects: effectively reducing wavelength-related spatial mismatch, improving edge spectral drift, avoiding damage to microstructures caused by non-physical elastic deformation, and reducing the smoothing of the true spectral shape by statistical equalization. It is particularly suitable for constructing large-field, high-fidelity hyperspectral microscopic panoramic images, providing a reliable data foundation for subsequent quantitative pathological analysis.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-fidelity stitching method for large-field hyperspectral microscopic images, characterized in that, Includes the following steps: S1. System-level physical calibration: Establishing wavelength-dependent spatial mapping relationships, full-spectrum energy attenuation models, and a unified absolute physical coordinate grid for panoramic stitching in an offline state; S2. Online acquisition and physical correction: Perform multi-field hyperspectral scanning on the target sample to obtain multiple overlapping local hyperspectral data cubes; perform physical correction on each local data cube and map it to the unified absolute physical coordinate grid to complete the initial global arrangement of each field of view; S3. Adaptive fusion of overlapping regions: Based on the spatial-spectral gradient continuity criterion, adaptive fusion processing is performed on the overlapping regions of adjacent fields of view; S4. Panoramic Data Construction: Integrate all fused field-of-view data into a unified absolute physical coordinate grid to generate a panoramic hyperspectral data cube.

2. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Wavelength-by-wavelength spatial distortion calibration: Using a calibration target with known absolute physical dimensions, calibration images are acquired at multiple consecutive wavelengths. For each wavelength, a mapping relationship or its inverse mapping relationship from the acquired pixel coordinate system to the physical coordinate system is established, forming a set of inverse mapping data indexed by wavelength. S12. Full-spectrum energy attenuation model calibration: Collect full-band dark field and bright field data, and calibrate the model for each wavelength. Construct the corresponding dark field image and bright field image The energy attenuation distribution function at this wavelength is obtained by smoothing and fitting the bright field image. By combining the attenuation distribution functions of all wavelengths, a full-spectrum energy attenuation model is formed; S13. Establishment of a unified absolute physical coordinate grid: Define the unified absolute physical coordinate grid based on the physical size of the calibrated target, the pixel size of the detector, the imaging magnification, and the imaging ratio of the system.

3. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 2, characterized in that, In S11, the calibration target is a high-precision grid target or a lattice target; the implementation of the mapping relationship or inverse mapping relationship includes radial distortion model, radial and tangential joint distortion model, polynomial geometric model, spline model and lookup table.

4. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 2, characterized in that, In S12, the smoothing fitting methods for the bright field image include two-dimensional polynomial fitting, low-rank surface fitting, and spline surface fitting; the full-spectrum energy attenuation model is used to perform radiometric correction on the original hyperspectral data, and the expression of the full-spectrum energy attenuation model is: .

5. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 1, characterized in that, S2 specifically includes the following: S21. Multi-field hyperspectral scanning acquisition: A hyperspectral microscopy imaging device and a motorized stage are used to scan the sample field by field, obtaining multiple overlapping hyperspectral local data cubes, and recording the stage position information corresponding to each field of view; the overlap rate between adjacent fields of view in the scanning direction is 10% to 30%; the scanning path includes line scanning, serpentine scanning, and block scanning; k The raw hyperspectral data obtained from each field of view are: ； S22. Radiometric Correction: Using the dark field data and full-spectrum energy attenuation model from S12, the corrected hyperspectral data is obtained. The correction method is wavelength-by-wavelength radiative compensation, calculated using the following formula: ； in, For raw hyperspectral data, This is the dark field image corresponding to wavelength λ. wavelength λ The corresponding full-spectrum energy decay model; A small positive constant set to prevent the denominator from being too small; S23. Wavelength-by-wavelength absolute coordinate restoration: For each field of view after radiometric correction, the corresponding inverse mapping data is called at each wavelength to reverse map the pixels from the acquisition coordinate system to the unified absolute physical coordinate grid, and the field of view is located to the initial position of the panoramic coordinate system in combination with the stage position information; the inverse mapping method includes bilinear interpolation and bicubic interpolation, which are used to obtain the pixel values ​​at non-integer coordinates. S24. Establishment of panoramic frame to be fused and identification of overlapping areas: Based on the initial position, field size and preset overlap relationship of each field of view determined in S23, establish the panoramic data frame to be fused and identify the overlapping areas between adjacent fields of view.

6. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 1, characterized in that, S3 specifically includes the following: S31. Overlapping Area Determination: Extract any two adjacent fields of view under a unified absolute physical coordinate system. A and B The overlapping region Ω; S32. Joint gradient feature construction: Calculate the features of two adjacent views within the overlapping region Ω. A and B The spatial gradient and spectral gradient are combined to form a joint spatial-spectral gradient feature; S33. Constructing the fusion cost function and solving for the optimal splicing boundary: Constructing the fusion cost function with the goal of minimizing the gradient abrupt change at the splicing seam, and solving for the optimal splicing boundary path through optimization methods; S34, Single-view pixel optimization writing: Divide the overlapping area into two sub-regions according to the optimal stitching boundary path, and each sub-region retains the pixels from the corresponding single view as the final output.

7. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 6, characterized in that, In S32, for any position The spatial gradient is defined as: ； The spectral gradient is defined as: ； The joint spatial-spectral gradient characteristics are: ； in, G xy ( x , y , λ () is the spatial gradient vector, describing the image at position. Local spatial structure changes; Image intensity I exist x The partial derivative in the direction reflects the rate of change in the horizontal direction; Image intensity I exist y The partial derivative in the direction reflects the rate of change in the vertical direction; G λ ( x , y , λ The spectral gradient represents the rate of change of light intensity in a hyperspectral image along the wavelength direction. Image intensity I For wavelength λ The partial derivatives; These are weighting coefficients used to balance the dimensional differences between spatial and spectral gradients.

8. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 7, characterized in that, In S33, the fusion cost function is: ； in: These are non-negative weighting coefficients; and Representing the field of view and vision In position Joint gradient at; and Representing the field of view and vision In position Image intensity value at; This is a confidence penalty term used to characterize the local noise level, edge signal-to-noise ratio, and penalty for distance from the view boundary; The boundary smoothing term is used to constrain the continuity of the spliced ​​boundary path and avoid excessively jagged boundaries; the optimization methods include dynamic programming, graph cut, shortest path search and graph optimization.

9. The high-fidelity stitching method for large-field hyperspectral microscopic images according to claim 8, characterized in that, In S34, the specific operation is as follows: the overlapping region is divided into sections belonging to the field of view. A and vision B Two sub-regions; for any overlapping region position p The final output pixel value is represented as: 。 10. A high-fidelity stitching system for large-field hyperspectral microscopic images, characterized in that, include: The system calibration module is used to establish wavelength-dependent spatial mapping relationships, full-spectrum energy attenuation models, and a unified absolute physical coordinate grid for panoramic stitching in an offline state. The online acquisition and correction module is used to perform multi-field hyperspectral scanning of the target sample to obtain multiple overlapping local hyperspectral data cubes. Each local data cube is physically corrected and mapped to the unified absolute physical coordinate grid to complete the initial global arrangement of each field of view. The overlapping region fusion module is used to adaptively fuse overlapping regions of adjacent fields of view based on the spatial-spectral gradient continuity criterion. The panorama construction module is used to integrate all fused field-of-view data into a unified absolute physical coordinate grid to generate a panorama hyperspectral data cube.

Images